What is graph-engineering?

This repository teaches a discipline called graph engineering, which is about designing the structures that AI agents work through rather than focusing on the words they use. The author splits it into two halves. Knowledge graphs are what agents remember: entities, facts, and relationships tracked with time and origin. Task graphs are how agents work: jobs connected by execution dependencies, with rules for parallel work, verification, and human approval checkpoints. The knowledge graph material comes from a graduate course at Southeast University, taught in Chinese since 2019. The repo translates and distills that course into English. It also incorporates modern research on agent orchestration from Google DeepMind, MIT, and Anthropic. The result is a set of practical tools rather than academic theory alone. Inside the repo you find a "skill" folder you can install into a compatible AI agent like Claude Code. Once installed, you can ask the agent to build a knowledge graph from your documents or to teach you the material interactively. The teaching mode walks through the pipeline stage by stage, using your own project as the running example with generated diagrams. A separate workflows file provides nine ready-to-use prompt blocks, including a tutor mode and eight single-purpose tools that chain together into a full build. The knowledge graph pipeline has nine stages: scope, representation, ontology, entities, relations, events, a quality gate, fusion, and serving to language models. The guiding principle is to model the domain before extracting information, fuse before storing, and verify at every stage. The author emphasizes that a knowledge graph is a product with a schema, not a pile of raw data. The task graph half offers specific rules for structuring agent work. Delete fake edges where no real work flows. Use a diamond pattern that splits work across parallel workers, runs verification in separate contexts, and merges through one owner. Research shows teams perform well on splittable work but lose on sequential work. A human approval gate should sit exactly where mistakes are costly to undo.